A three-dimensional Gaussian reconstruction method for underwater scene based on double-branch enhanced constraint

Abstract

The application discloses a kind of underwater scene three-dimensional Gaussian reconstruction methods based on double branch enhancement constraint, it is related to underwater visual perception and three-dimensional reconstruction technical field, after frame screening, underwater multi-view image is divided into pose branch and appearance branch and handles in parallel.Pose branch uses near original map conservative strategy, and COLMAP pose solution is guaranteed stable;Appearance branch is enhanced in view of underwater degradation, and rendering quality is improved.Through consistency alignment module, double branch data binding is completed, input WaterSplatting realizes three-dimensional Gaussian primitive reconstruction and optimization, and finally multi-format three-dimensional result is output.The application uses the above method, decouples the input constraint of geometry recovery and appearance reconstruction, considers pose accuracy and visual effect, and is suitable for a variety of underwater scene digitization application.

Description

Technical Field

[0001] This invention relates to the field of underwater visual perception and 3D reconstruction technology, and in particular to a method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints. Background Technology

[0002] Underwater 3D reconstruction is a key supporting technology for marine resource surveys, underwater archaeology, and marine ranching monitoring. In recent years, neural rendering methods, represented by Neural Radiation Field (NeRF) and 3D Gaussian Sputtering (3DGS), have made significant progress in 3D reconstruction of terrestrial scenes and are gradually being applied to underwater scenes. Among them, WaterSplatting extends the 3D Gaussian sputtering framework to the underwater environment, while SeaThru-NeRF introduces a scattering medium physical model within the NeRF framework and provides public datasets, laying the foundation for quantitative evaluation of underwater 3D reconstruction.

[0003] However, during underwater image acquisition, light propagation is affected by multiple physical factors such as selective wavelength absorption, scattering, and occlusion by suspended particles, leading to degradation issues in images, including a bluish-green tint, low contrast, blurred textures, and inconsistent appearance across frames. In the underwater 3D reconstruction process, feature-matching-based structure-of-motion (SfM) algorithms such as COLMAP require good geometric consistency and appearance matching between adjacent frames, while neural rendering methods such as WaterSplatting require excellent color representation and texture visibility in the input image. These two requirements are clearly contradictory and coupled, representing a core challenge restricting the improvement of reconstruction quality.

[0004] Existing underwater 3D reconstruction schemes mainly employ two approaches for image preprocessing: one is the no-enhancement approach, which directly uses the original underwater image to drive reconstruction. While this ensures consistency across frames, it results in sparse feature points and poor rendering quality in low-visibility scenes. The other is the uniform enhancement approach, which applies a uniform strong enhancement to the image before reconstruction. Although this improves the visual effect of a single frame, it disrupts cross-frame consistency, leading to a degradation in pose estimation accuracy. The core flaw of existing technologies is that they share the same set of enhancement results with the geometric pose recovery and appearance rendering chains. This makes it impossible to simultaneously meet the differentiated requirements of the two chains for the input image, and it is difficult to balance pose estimation stability and rendering quality. Summary of the Invention

[0005] The purpose of this invention is to provide a method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints, which explicitly decouples the geometric pose recovery link and the appearance neural rendering link from the coupling mode of sharing the same set of enhancement images, fundamentally eliminating the inherent contradiction of mutual constraints between the two links caused by the unified enhancement strategy.

[0006] To achieve the above objectives, this invention provides a method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints, comprising the following steps: S1. Acquire multi-view image sequences of the target underwater scene and perform frame quality filtering; S2. Construct pose branch images from the filtered images; S3. Construct an appearance branch image from the images filtered in step S2; S4. Using the pose branch image as input to COLMAP, feature matching, camera intrinsic and extrinsic parameter solving, sparse point cloud reconstruction and image distortion removal are completed, and the standard format sparse point cloud model, distorted image and corresponding pose information are output. S5. Through the consistency alignment module, the appearance branch image is integrated with the corresponding pose information and sparse point cloud model output in step S4. S6. Input the aligned appearance branch image and pose information into WaterSplatting or a compatible 3D Gaussian primitive reconstructor. S7. Export reconstruction results.

[0007] Preferably, step S1 contains the following: Based on Laplacian variance scoring to remove blurry frames, the sharpness scoring formula is as follows: ; in, For the Laplacian response of the image; For variance operators; when Time frame preservation, Time frame culling; among which For sharpness threshold; Remove near-duplicate frames based on perceptual hash similarity.

[0008] Preferably, step S2 contains the following: A light gray-world white balance correction was performed on the selected underwater image sequences, with the correction intensity set to no more than 0.3; A weak red channel compensation is performed on the corrected image, with the red channel gain set to no more than 1.2; Perform low-intensity CLAHE contrast adjustment on the compensated image, wherein CLAHE is set to off, or the clipLimit parameter is set to no more than 1.5; Output pose branch images, which maintain consistent appearance across frames and have no obvious local histogram transformations or stylization changes, for use in subsequent camera pose solving.

[0009] Preferably, step S3 is as follows: The red channel adaptive compensation, gray world white balance, mild saturation enhancement, and optional limited contrast adaptive histogram equalization (CLAHE) are performed sequentially to enhance texture visibility and color expression. Each enhancement parameter remains fixed within the same image sequence to avoid introducing cross-frame brightness style jumps due to frame-by-frame adaptive adjustments.

[0010] The preferred adaptive compensation formula for the red channel is as follows: ; in, The pixel value of the c-th channel of the original image. The water attenuation coefficient for the corresponding channel ( (Red light attenuation is strongest), where d is the normalized water depth estimate. To compensate for the pixel values, For the red channel, For green channel, The blue channel.

[0011] Preferably, in step S5, the image file name, frame index order and camera pose entry are strictly matched one-to-one to eliminate the data arrangement inconsistency problem that may be introduced by the parallel processing of dual branches, and form a standard input data structure that WaterSplatting can directly read.

[0012] Preferably, step S6 contains the following: Initialize a 3D Gaussian primitive set using a COLMAP sparse point cloud; The position, covariance, opacity, and spherical harmonic color parameters of each Gaussian element are iteratively optimized using differentiable rasterization rendering and a joint loss function, and adaptive densification and pruning control are performed. The joint loss function is as follows: ; in, For pixel-level L1 reconstruction loss, The loss function is defined as structural similarity loss, with λ representing the weighting coefficients. This loss function jointly constrains the optimization process of the 3D Gaussian primitives in two dimensions: pixel accuracy and perceptual structure.

[0013] Preferably, the reconstruction results exported in step S7 include rendered images from each test viewpoint, sparse point clouds in PLY format, 3D meshes in PLY format, and meshes in OBJ format, which can be directly used for underwater scene digital archiving, visualization, subsequent fine-tuning modeling, or other downstream applications.

[0014] Therefore, the present invention employs the above-mentioned method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhanced constraints, which has the following beneficial effects: (1) The proposed dual-branch explicit decoupling architecture decomposes the input underwater image sequence into pose branch images and appearance branch images in parallel, allowing COLMAP camera pose recovery and WaterSplatting 3D Gaussian sputtering rendering to use different input images respectively, fundamentally removing the coupling constraints of geometric recovery and appearance reconstruction on the input images. This architecture not only ensures the stability of camera pose solution but also improves the visual quality of the rendering output, achieving synergistic optimization between the two.

[0015] (2) The pose branch adopts a near-original image conservative processing strategy, applying only mild gray-world white balance or slight color cast compensation, disabling local histogram enhancement or setting the contrast limiting factor to an extremely low value, thus maximizing the protection of cross-frame feature appearance consistency. This strategy avoids interference from image enhancement operations on feature matching, significantly improves the feature matching rate and camera registration stability of COLMAP in underwater scenes, and provides a reliable geometric basis for subsequent 3D reconstruction.

[0016] (3) The appearance branch, in response to underwater optical degradation characteristics, sequentially performs a combination of operations such as red channel adaptive compensation, gray world white balance, mild saturation enhancement, and optional CLAHE, effectively compensating for the selective absorption of long-wavelength light by water and improving the color reproduction and texture detail visibility of the image. At the same time, by fixing the enhancement parameters within the same image sequence, the cross-frame brightness style jumps introduced by frame-by-frame adaptive adjustment are avoided, providing high-quality, style-consistent input data for 3D Gaussian sputtering rendering and significantly improving the visual expressiveness of the reconstruction results.

[0017] (4) A dedicated consistency alignment module was designed to strictly align the camera intrinsic and extrinsic parameters, sparse point cloud and appearance branch image output by the pose branch according to the image file name and frame index order, forming a standard input data structure that WaterSplatting can directly read. This effectively eliminates the data arrangement inconsistency problem that may be introduced by parallel processing of dual branches, and ensures the continuity and reliability of the entire reconstruction process.

[0018] (5) The dual-branch preprocessing module, COLMAP motion recovery structure algorithm, WaterSplatting 3D Gaussian primitive reconstructor and consistency alignment module are integrated into a unified end-to-end underwater 3D reconstruction system. The process is highly automated and can directly output 3D results in multiple formats such as rendered images, PLY point clouds, PLY meshes and OBJ meshes. It can meet the needs of various downstream applications such as underwater scene digital archiving, visualization, and subsequent fine modeling.

[0019] (6) It supports independent parameter adjustment of the pose branch and appearance branch based on the color cast of the target scene, water visibility, and cross-frame matching stability, which has strong adaptability. Experimental verification shows that the parameters optimized for specific scenes can be directly transferred to other underwater scenes and maintain significant performance improvement. It has practical cross-scene generalization ability and reduces the parameter tuning cost in practical applications.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of a method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhanced constraints according to the present invention. Figure 2 This is a schematic diagram of the dual-branch enhancement and reconstruction structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the comparative experimental grouping in an embodiment of the present invention; Figure 4 This is a comparison of the preprocessing results of the Panama scene in the embodiments of the present invention; wherein, (a) is the original image; (b) is a schematic diagram of the pose branch light enhancement result: used for COLMAP pose recovery and sparse reconstruction; (c) is a schematic diagram of the appearance branch enhancement result: used for WaterSplatting appearance reconstruction; Figure 5 This is a schematic diagram comparing the COLMAP sparse reconstruction results of the Panama scene in an embodiment of the present invention; Figure 6 The following is a comparison of the rendering results of various methods in the Panama scene according to embodiments of the present invention; wherein, (a) is the original reference image; (b) is a schematic diagram of the rendering result obtained by directly performing WaterSplatting reconstruction (Baseline B) with the original image and COLMAP pose; (c) is a schematic diagram of the rendering result obtained by replacing the original COLMAP pose with an enhanced image for WaterSplatting reconstruction (Ablation D); and (d) is a schematic diagram of the rendering result obtained by reconstructing using the dual-branch enhancement constraint method of the present invention. Figure 7 This is a schematic diagram comparing the rendering results of various methods in the Curasao scene according to an embodiment of the present invention; wherein, (a) is the original reference image; (b) is a schematic diagram of the rendering result obtained by directly performing WaterSplatting reconstruction using the original image and COLMAP pose; (c) is a schematic diagram of the rendering result obtained by replacing the original COLMAP pose with an enhanced image for WaterSplatting reconstruction while keeping the original COLMAP pose unchanged; and (d) is a schematic diagram of the rendering result obtained by reconstructing using the dual-branch enhancement constraint method of the present invention. Figure 8 The following is a comparison of the rendering results of various methods for the Japanese Gardens-RedSea scene in the embodiments of the present invention; wherein, (a) is the original reference image; (b) is a schematic diagram of the rendering result obtained by directly performing WaterSplatting reconstruction using the original image and COLMAP pose; (c) is a schematic diagram of the rendering result obtained by replacing the original COLMAP pose with an enhanced image for WaterSplatting reconstruction while keeping the original COLMAP pose unchanged; and (d) is a schematic diagram of the rendering result obtained by reconstructing using the dual-branch enhancement constraint method of the present invention. Figure 9 The following is a comparison of the rendering results of various methods in the IUI3-RedSea scene according to the embodiments of the present invention; wherein, (a) is the original reference image; (b) is a schematic diagram of the rendering result obtained by directly performing WaterSplatting reconstruction using the original image and COLMAP pose; (c) is a schematic diagram of the rendering result obtained by replacing the original COLMAP pose with an enhanced image for WaterSplatting reconstruction while keeping the original COLMAP pose unchanged; and (d) is a schematic diagram of the rendering result obtained by reconstructing using the dual-branch enhancement constraint method of the present invention. Figure 10 This is a comparison of rendering quality before and after Panama scene optimization in an embodiment of the present invention; wherein, (a) is the original reference image; (b) is the Baseline B rendering result: a schematic diagram of the rendering result corresponding to the standard baseline scheme; (c) the optimized reference image; and (d) the rendering result of the present invention after optimization: a schematic diagram of the rendering result after using the optimal parameter configuration of the Panama scene.

[0022] Figure 11 This is a schematic diagram of the Panama scene migration result with optimal Panama parameters according to an embodiment of the present invention; wherein, (a) is the original Panama reference image; (b) is the Panama scene Baseline B rendering result; (c) is the Panama optimization reference image; and (d) is the rendering result of the present invention after Panama optimization. Figure 12 This is a schematic diagram of the scene migration result of Curasao with optimal Panama parameters according to an embodiment of the present invention; wherein, (a) is the original reference image of Curasao; (b) is the Baseline B rendering result of the Curasao scene; (c) is the Curasao optimization reference image; and (d) is the rendering result of the present invention after Curasao optimization. Figure 13This is a schematic diagram of the scene migration results of JapaneseGradens-RedSea with optimal Panama parameters in an embodiment of the present invention; wherein, (a) is the original reference image of JapaneseGradens-RedSea; (b) is the Baseline B rendering result of JapaneseGradens-RedSea scene; (c) is the optimized reference image of JapaneseGradens-RedSea; and (d) is the rendering result of the present invention after the optimization of JapaneseGradens-RedSea. Figure 14 This is a schematic diagram of the Panama optimal parameter IUI3-RedSea scene migration result in an embodiment of the present invention; wherein, (a) is the original reference image of IUI3-RedSea; (b) is the IUI3-RedSea scene Baseline B rendering result; (c) is the IUI3-RedSea optimization reference image; and (d) is the rendering result of the present invention after IUI3-RedSea optimization. Figure 15 This is a schematic diagram of the export results of the 3D point cloud and mesh of the Panama scene according to an embodiment of the present invention, wherein (a) is a schematic diagram of the export results of the 3D point cloud of the Panama scene; and (b) is a schematic diagram of the export results of the mesh. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example Please see Figures 1-2This invention provides a method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints. The pose branch performs only mild gray-world white balance or slight color shift compensation close to the original image, disabling local histogram enhancement or setting its contrast limiting factor to an extremely low value to maximize the matching compatibility of cross-frame features. The appearance branch performs a combination of red channel adaptive compensation, gray-world white balance, mild saturation adjustment, and optional local contrast enhancement. The steps are as follows: Step 1: Acquire a multi-view image sequence of the target underwater scene and perform frame quality screening. Blurry frames are removed based on Laplacian variance scoring, and nearly duplicate frames are removed based on perceptual hash similarity to improve the input quality and stability of subsequent reconstruction. The sharpness scoring formula is as follows: ; in, For the Laplacian response of the image; For variance operator; when Time frame preservation, Time frame culling; among which This is the sharpness threshold. The lower the variance, the blurrier the image; frames below the threshold will be discarded.

[0026] Step 2: Construct pose branch images for the selected images. This branch aims to protect the consistency of appearance across frames, performing only mild white balance or slight color cast compensation to avoid introducing local histogram transformations or high-intensity contrast adjustments, thus ensuring that the stability of COLMAP feature matching is not disturbed.

[0027] Step 3: Construct an appearance branch image from the same batch of selected images. This branch aims to improve the input quality of neural rendering, sequentially performing red channel adaptive compensation (compensating for the selective absorption of long-wavelength light by water), gray-world white balance, mild saturation enhancement, and optional limited contrast adaptive histogram equalization (CLAHE) to enhance texture visibility and color expression. Each enhancement parameter remains fixed within the same image sequence to avoid frame-by-frame adaptive adjustments that introduce cross-frame brightness style jumps. The red channel compensation formula is as follows: ; in, The pixel value of the c-th channel of the original image. The water attenuation coefficient for the corresponding channel ( (Red light attenuation is strongest), where d is the normalized water depth estimate. The pixel value after compensation.

[0028] Step 4: Using the pose branch image as COLMAP input, perform feature extraction, exhaustive feature matching and incremental structure of motion restoration (SfM) algorithm to recover the sparse 3D point cloud involving the camera inside and outside each frame, and complete the distortion removal process through the image_undistorter module to output the sparse model and distortion-removed image in standard COLMAP format.

[0029] Step 5: Through the consistency alignment module, the appearance branch images are uniformly integrated with the corresponding pose information and sparse point cloud model output from Step 4. This ensures that the image file names, frame index order, and camera pose entries strictly correspond one-to-one, eliminating data arrangement inconsistencies that may be introduced by parallel processing of dual branches, and forming a standard input data structure that WaterSplatting can directly read.

[0030] Step 6: Input the aligned appearance branch image and pose information into WaterSplatting or a compatible 3D Gaussian primitive reconstructor. Initialize the 3D Gaussian primitive set with COLMAP sparse point cloud. Iteratively optimize the position, covariance, opacity, and spherical harmonic color parameters of each Gaussian primitive through differentiable rasterization rendering and a joint loss function. Perform adaptive densification and pruning control to finally obtain a high-quality rendering result and a 3D Gaussian field model. The joint loss function is as follows: ; in For pixel-level L1 reconstruction loss, The loss function is defined as structural similarity loss, with λ representing the weighting coefficients. This loss function jointly constrains the optimization process of the 3D Gaussian primitives in two dimensions: pixel accuracy and perceptual structure.

[0031] Step 7: Export the reconstruction results, including rendered images from each test viewpoint, sparse point clouds in PLY format, 3D meshes in PLY format, and meshes in OBJ format. These can be directly used for underwater scene digital archiving, visualization, subsequent fine-tuning modeling, or other downstream applications.

[0032] like Figures 3-15 In the optimization implementation of the Panama scene, the pose branch adopted a parameter configuration close to the original image (pose preservation series), while the appearance branch adopted mild color restoration and disabled CLAHE (pose preservation_rendering_equalization configuration). Experimental results show that this configuration maintains the stability of COLMAP sparse reconstruction (the number of registered images and reprojection error are both on par with the baseline of the original image) while achieving the best results in the WaterSplatting quantitative index. Furthermore, the optimal parameters of the Panama scene were directly transferred to the Curasao, JapaneseGardens-RedSea, and IUI3-RedSea scenes for cross-scene re-runs, verifying that the parameter configuration has good scene generalization ability.

[0033] The evaluation results of WaterSplatting in the standard unified parameter experiments of four public underwater scenarios are shown in Table 1.

[0034] Table 1. Evaluation Results of Standard Configuration Experiments in Four Public Scenarios

[0035] In the Panama scene, after further optimization of the dual-branch parameters of the present invention, the WaterSplatting evaluation results are shown in Table 2. The scheme of keeping the pose branch close to the original image and applying slight color restoration to the appearance branch achieved the best results.

[0036] Table 2 Evaluation Results of Panama Scene Optimization Examples

[0037] Table 3 shows the COLMAP statistics for different pose branch configurations in the Panama scene. When the pose branch is close to the original image, the reprojection error is closer to the result of the original image.

[0038] Table 3. COLMAP Statistical Results of Pose Branches in Panama Scene

[0039] The optimal pose-preserving rendering-equalization configuration of the Panama scene was directly transferred to the other three public scenes, resulting in cross-scene reruns, as shown in Table 4. This section verifies the portability of the parameter design of this invention and demonstrates that the optimized dual-branch scheme has a significant stability improvement compared to the initial standard configuration.

[0040] Table 4. Comparison results of the present invention with the baseline after optimization in four scenarios.

[0041] The average results across four scenarios show that: Baseline B has average peak signal-to-noise ratio (PSNR), structural similarity, and learned perceptual image patch similarity of 29.3425 / 0.929227 / 0.135045, respectively; the initial standard configuration of this invention has average PSNR, structural similarity, and learned perceptual image patch similarity of 23.6390 / 0.836264 / 0.226191, respectively; and after optimization, the average PSNR, structural similarity, and learned perceptual image patch similarity of this invention are 29.2353 / 0.928822 / 0.137114, respectively.

[0042] Therefore, this invention adopts the above-mentioned method for 3D Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints, which prioritizes the pose recovery link to ensure cross-frame appearance consistency and feature matching stability, and prioritizes the appearance rendering link to improve the texture visibility and color expression quality of the input image, thereby giving full play to the detail restoration capability of the 3D Gaussian sputtering neural rendering framework without sacrificing pose accuracy.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for 3D Gaussian reconstruction of underwater scenes based on dual-branch augmentation constraints, characterized in that, Includes the following steps: S1. Acquire multi-view image sequences of the target underwater scene and perform frame quality filtering; S2. Construct pose branch images from the filtered images; S3. Construct an appearance branch image from the images filtered in step S2; S4. Using the pose branch image as input to COLMAP, feature matching, camera intrinsic and extrinsic parameter solving, sparse point cloud reconstruction and image distortion removal are completed, and the standard format sparse point cloud model, distorted image and corresponding pose information are output. S5. Through the consistency alignment module, the appearance branch image is integrated with the corresponding pose information and sparse point cloud model output in step S4. S6. Input the aligned appearance branch image and pose information into WaterSplatting or a compatible 3D Gaussian primitive reconstructor. S7. Export reconstruction results.

2. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 1, characterized in that, Step S1 is as follows: Based on Laplacian variance scoring to remove blurry frames, the sharpness scoring formula is as follows: ； in, For the Laplacian response of the image; For variance operators; when Time frame preservation, Time frame culling; among which For sharpness threshold; Remove near-duplicate frames based on perceptual hash similarity.

3. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 2, characterized in that, Step S2 is as follows: A light gray-world white balance correction was performed on the selected underwater image sequences, with the correction intensity set to no more than 0.3; A weak red channel compensation is performed on the corrected image, with the red channel gain set to no more than 1.2; Perform low-intensity CLAHE contrast adjustment on the compensated image, with CLAHE set to off, or set the clipLimit parameter to no more than 1.5; The output pose branch image maintains consistent appearance across frames, with no obvious local histogram transformations or stylization changes, and is used for subsequent camera pose solving.

4. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 3, characterized in that, Step S3 is as follows: In sequence, red channel adaptive compensation, gray world white balance, mild saturation enhancement, and contrast-limited adaptive histogram equalization (CLAHE) are performed to enhance texture visibility and color expression. Each enhancement parameter is fixed within the same image sequence.

5. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 4, characterized in that, The adaptive compensation formula for the red channel is as follows: ； in, The pixel value of the c-th channel of the original image. This represents the water attenuation coefficient for the corresponding channel. This is a normalized water depth estimate. To compensate for the pixel values, For the red channel, For green channel, The blue channel.

6. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 5, characterized in that: In step S5, the image file name and frame index order are matched one-to-one with the camera pose entries to eliminate the data arrangement inconsistency problem caused by the parallel processing of the two branches, forming a standard input data structure that WaterSplatting can directly read.

7. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 6, characterized in that, Step S6 is as follows: Initialize a 3D Gaussian primitive set using a COLMAP sparse point cloud; The position, covariance, opacity, and spherical harmonic color parameters of each Gaussian element are iteratively optimized using differentiable rasterization rendering and a joint loss function, and adaptive densification and pruning control are performed. The joint loss function is as follows: ； in, For pixel-level L1 reconstruction loss, λ represents the structural similarity loss, and λ is the weighting coefficient.

8. The method for three-dimensional Gaussian reconstruction of underwater scenes based on dual-branch enhancement constraints according to claim 7, characterized in that: The reconstruction results exported in step S7 include rendered images from each test viewpoint, sparse point clouds in PLY format, 3D meshes in PLY format, and meshes in OBJ format.

Images